> I expect the problem isn’t necessarily the AI generation itself, but the difficulty to distinguish between someone that’s really understood the problem domain versus someone with LLM-induced “Dutch courage”.

Some of that. Also sometimes there isn't even a person, just a bot opening a pull request that doesn't hear feedback.

There's also the problem that many projects put disproportionate effort into mentoring new contributors and helping them get things right, and that's valuable when there's a human learning something, but not when you're just correcting someone's LLM slop.

And that's leaving aside all the other problems, like "these things took all our FOSS code, ignoring its licenses, and then used it to compete with us and drown us out", or "these things are going to turn us all into paperclips".

Personally, I'd say the right way to use an LLM is to let it review your code, or be your rubber duck helping you design the architecture or evaluate the security model or similar, but not let it write the code for you. (Unless it's truly throw-away code, like "write me some code to benchmark X vs Y to evaluate which is better for my use case". In which case, the point wasn't the code, it was the numbers.)